Title: Now-casting financial volatility with long memory: A non-Gaussian and non-linear state space approach
Authors: Yuze Liu - University of Cologne (Germany) [presenting]
Abstract: Recently, a simple model for now-casting daily financial log-volatility has been proposed. In contrast to existing approaches, it uses current and past information. It obeys an ARMA representation for log-squared returns and is related to the well-known SV model. This model outperforms the EGARCH model and the SV model. However, there are some important limitations. First, the normality assumption on daily returns is critical. Second, the ML estimation under a Gaussian approximation is biased and inefficient in finite-samples. Third, the typical long memory feature is not captured. These issues are tackled by considering a flexible non-Gaussian and non-linear ARMA representation of the log-transformed squared returns. The strong dependence and leverage in volatility are captured by an asymmetric ARFIMA model. It is estimated via the suitable Kitagawa state-space filter. It implements the numerical exact ML estimation under non-Gaussian and non-linear distributions. The employed estimation framework is flexible enough to cover the non-normality of daily returns explicitly. In an extensive Monte Carlo study, bias-reduction and efficiency gains are investigated. The volatility now-casting performance is evaluated by means of MSE and QLIKE. In an empirical application, volatility connectedness of US bond markets is studied in comparison to the commonly applied GARCH(1,1) model.